cbert33/DeepSeek-V4-Flash-0731-abliterated-vision

DeepSeek-V4-Flash-0731 Abliterated Vision

We were happy with Cebeuq's DeepSeek-V4-Flash-0731 abliterated checkpoint, but we wanted vision too. We did not want to spend roughly another 8 GB on the Kimi-derived vision towers used by the other DeepSeek vision grafts we considered, so we built this smaller composition ourselves from the Cebeuq language model and FlyCockpit's DeepEncoderV2 tower and projector.

This model requires our companion dgx-spark-vllm-deepseek-v4-vision runtime repository to use its vision features. Stock Transformers and stock vLLM will not load the custom multimodal architecture correctly; use the pinned Anemll-based runtime and deployment instructions from that repository.

Uncensored model: the language checkpoint has undergone abliteration to reduce refusal behavior. Treat outputs as untrusted, apply application-level safeguards, and do not assume the model will decline harmful requests.

User responsibility: this model is provided without warranty. The maintainers are not responsible for what others generate, publish, deploy, or otherwise do with this abliterated model. Users must operate it responsibly, apply appropriate safeguards, comply with applicable law, and respect third-party rights.

What “merged” means here

This is a composition, not weight averaging or parameter blending:

1. The complete language-model repository is copied from cebeuq/DeepSeek-V4-Flash-0731-abliterated at revision 21bd923c2574d9edcd7b914885024ce72fd5c076.[2]
2. The DeepEncoderV2 tower and trained projector are copied from FlyCockpit/DeepSeek-V4-Flash-0731-vision at revision d8efc7dfaceee965164d95952e2498b60fee323c.[3]
3. config.json is updated to declare DeepseekV4VisionForCausalLM and record the image-token and preprocessing contract.
4. The original model.safetensors.index.json is required to remain byte-for-byte unchanged during assembly. No language tensor is rewritten by the vision composition step.
5. At inference, each image is encoded by DeepEncoderV2, projected from 896 dimensions into the language model’s 4096-dimensional embedding space, and spliced into the prompt at image token ID 129279. A learned separator follows the view embeddings.

The underlying abliterated checkpoint itself is derived from DeepSeek’s official DeepSeek-V4-Flash-0731 release.[1][2]

Pinned components

Component	Source	Revision / digest
Language checkpoint	cebeuq/DeepSeek-V4-Flash-0731-abliterated	revision 21bd923c2574d9edcd7b914885024ce72fd5c076; weight-index SHA-256 a93ace48e04f89e17222353191d30a5ba5744cbb8176a7a4832221656ce2d545; 50-entry weight-manifest SHA-256 6ddeea70678d7f11a6ea72514cbb799b7070617e9fa3d62f990722e27ba371f4
Declared official base	deepseek-ai/DeepSeek-V4-Flash-0731	the Cebeuq card does not declare the exact official-base revision used
Vision tower + projector	FlyCockpit/DeepSeek-V4-Flash-0731-vision	d8efc7dfaceee965164d95952e2498b60fee323c
Tower file	vision/tower/deepencoder_v2_tower.safetensors	SHA-256 9dcf6803d4c6b63acc4008bc2409e599a2ab6e3886e241f1727f61550c300df5
Projector file	vision/adapter/merged-004800-5af0c5.pt	SHA-256 6d0235333941210666bf347abb95e334943ef3f230dac65b83551186925468ec
Runtime base	Locally supplied digest-pinned Anemll DGX Spark image	registry digest sha256:a83948492cf13df455170fb42885f5ef4db54fefe0feff0f841ecbff464ac9d8
Runtime source	Anemll GitHub repository ID 1301198905	47503f8e38dadd4dededca798150db2619594fce[4]
Vision plugin lineage	FlyCockpit/DeepSeek-V4-Vision-2x-DGX-Sparks	7cb20472e0f007a0626bd22ed9f5e22a8825c7e1[5]

The machine-readable copy of these pins is in SOURCE_PINS.json. MANIFEST.sha256 covers every other file in the assembled repository; the manifest does not list itself.

Vision contract

• Architecture: DeepEncoderV2 tower + MLP projector + learned view separator
• Input image size: 1024 × 1024
• Image token: <｜image｜> (129279)
• Tokens per view: 256
• Tiling: up to four local crops from a 2×2 grid, with one global view first and row-major crop order
• Image token counts: 257, 769, or 1281, depending on aspect-aware tiling
• Main language KV cache format in the qualified runtime: nvfp4_ds_mla

Qualified runtime

The release was exercised with:

• two NVIDIA GB10 systems in tensor parallel (TP=2);
• Anemll DSpark vLLM image 0.1.1 plus the included vision plugin;
• served model name DeepSeek-V4-Flash-0731-Vision;
• max_model_len=800000, max_num_seqs=2, and gpu_memory_utilization=0.88;
• measured startup cache admission of 1,791,777 tokens, sufficient for two full 800K requests with 191,777 aggregate tokens of admission headroom.

The measured number is specific to this exact hardware/runtime/profile. Re-profile after changing the GPU, runtime, graph mode, batching limits, vision implementation, or cache format.

Serving

Use the required dgx-spark-vllm-deepseek-v4-vision runtime repository and point it at this full model directory. The essential launch settings are:

--served-model-name DeepSeek-V4-Flash-0731-Vision
--tensor-parallel-size 2
--kv-cache-dtype nvfp4_ds_mla
--block-size 256
--max-model-len 800000
--max-num-seqs 2
--gpu-memory-utilization 0.88

Example OpenAI-compatible request:

curl http://HEAD:8304/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "DeepSeek-V4-Flash-0731-Vision",
    "chat_template_kwargs": {"thinking": false},
    "messages": [{"role": "user", "content": [
      {"type": "text", "text": "Read the text in this image."},
      {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
    ]}],
    "max_tokens": 128
  }'

Validation performed

The qualified two-node deployment passed:

• API health and model-list checks;
• a deterministic text response;
• two independent image/OCR prompts (ORBIT-7391 and NOVA-2846);
• rank health checks with no OOM;
• verification that the native DSpark/EAGLE3 path and nvfp4_ds_mla cache remained active.

This is a bounded deployment smoke test, not a comprehensive vision benchmark or safety evaluation. The redacted artifact identities, profile, observed outputs, and limitations are recorded in VALIDATION.md.

Limitations

• The vision adapter is primarily described and tested for screenshots and UI understanding; it is not established as a production computer-use or coordinate-grounding model.[3][5]
• Only one image per prompt was qualified.
• Stock Transformers and stock vLLM are unsupported for this assembled layout.
• The language component is abliterated/uncensored and may generate unsafe, misleading, or policy-violating content.[2]
• OCR smoke tests do not establish broad visual reasoning quality.
• The 800K × 2 result is an admission/cache measurement, not proof that every possible pair of 800K prompts will complete under every workload.

Licensing and redistribution

This assembled repository uses Apache-2.0 as its top-level license. The incorporated MIT notices are retained under LICENSES/, and third-party components retain their applicable upstream terms:

• DeepSeek’s base model and the abliterated language checkpoint identify MIT terms.[1][2]
• The FlyCockpit model repository is tagged Apache-2.0 and its card identifies the adapter packaging as Apache-2.0.[3]
• The same card says the DeepEncoderV2 tower remains subject to upstream terms.[3]
• The runtime overlay is derived from MIT-licensed FlyCockpit packaging and runs over Anemll’s runtime; vLLM-derived code in Anemll remains Apache-2.0.[4][5]

This top-level license choice does not remove or replace retained third-party notices. See THIRD_PARTY_NOTICES.md.

Reproducibility

See ASSEMBLY.md for the exact composition algorithm and verification gates. The release assembler:

• rejects symlinked or overlapping input trees;
• validates all 48 language shards, the overlay shard, and the language-model weight index against the pinned 50-entry LANGUAGE_WEIGHTS.sha256 manifest;
• requires every weight-index reference to be present in that manifest;
• validates pinned sizes and SHA-256 hashes for the tower, projector, and runtime wheel;
• copies into a staging directory and atomically renames it;
• rejects any change to the language-model weight index;
• emits a complete SHA-256 manifest.

Citation

If you use this package, cite the original DeepSeek V4 release and credit both component repositories:

@misc{deepseekai2026deepseekv4,
  title  = {DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
  author = {DeepSeek-AI},
  year   = {2026}
}

Also cite the exact pinned source URLs below so the composition can be reconstructed.

Sources

[1] DeepSeek-V4-Flash-0731 [2] DeepSeek V4 Flash 0731 Abliterated [3] DeepSeek V4 Flash 0731 Vision Assets [4] Anemll DGX Spark vLLM source [5] FlyCockpit DeepSeek V4 Vision Runtime